was able to create polygons of individual panel areas, albeit limited by imperfect linearity.
The second core technology is one enabling automatic diagnosis of defective panels based on extracted PV panel areas. Tsanakas
et al. 2015 designed a method to identify the location of hot spot cells on a PV panel using the Canny edge operator. In the
PV power plant maintenance and repair regime generally applied in South Korea, any panel containing defective cells is
replaced in its entirety. Hence, the algorithm should focus on diagnosing the function or malfunction of each panel rather than
identifying the locations of hot spot cells within a panel. Consequently, it is considered reasonable to employ a method
for comparing the intensity characteristics of individual panel area polygons, rather than one using threshold values such as
the Canny edge operator, for developing an algorithm in this study. In a related previous study, Kim et al. 2016a proposed
using statistical characteristics of thermal intensity as parameters for panel fault diagnosis because defective panels
display different patterns of intensity compared to intact panels. Drawing on this finding, this study aims to develop an
algorithm capable of automated PV panel fault diagnosis using intensity-related statistical values of each panel based on
extracted panel area polygons.
2. ANALYSIS OF PV PANEL INTENSITY
CHARACTERISTICS IN A THERMAL INFRARED IMAGE
As experimental data, we used a sample image Figure 1 a obtained from a UAV with FLIR T620 thermal imaging camera,
provided online by paul kitawa at a pixel size of 640×480. In a previous study Kim et al., 2016b, we developed an image
segmentation-based panel area extraction algorithm and created polygons for individual panel areas as shown in Figure 1 b.
The study confirmed that panel areas could be expressed with 93.9 accuracy compared with visually extracted and manually
digitized panel boundaries using the performance assessment method of McGlone et al. 2004. In Figure 1 c, each panel is
labelled with a serial number assigned to each extracted panel area polygon. Visual inspection of the labelled panels reveals
that hot spots are present on a total of five panels 6, 23, 26, 64, and 66 a string form. Analysis of the thermal characteristics of
defective PV modules suggests these hot spots are due to defective bypass diodes within the modules.
To reduce fault diagnosis error, it is essential to consider the overall intensity characteristics of each panel. We compared
intensity histograms in order to determine the intensity characteristics of normal and defective panels Figure 2.
Figure 1. a Sample thermal infrared image www.kiwata.de, b Panel area extraction result, and c Grayscale image with
PV module numbers labelled Figure 2. Examples of intensity histogram and statistical
characteristics of normal and defective panels The histograms on the left side in Figure 2 represent normal
panels 25 and 63, and those on the right side represent defective panels 26 and 64. Because intensity distributions of
normal panels are similar to those of adjacent panels, normal panels close to panels with hot spots were selected for
comparison. The blue solid and dotted lines in Figure 2 represent the mean intensity and standard deviation range,
respectively, of pixels in the entire panel area. The red diamond and solid line represent the mean intensity and standard
deviation range, respectively, of the pixels in each panel area concerned. The normal panel histograms are distributed at an
intensity range of 200 or lower and display patterns similar to the normal distribution of the entire panel area. In contrast, the
defective panel histograms contain pixels with intensity values exceeding 200 and display multiple peaks that deviate from a
normal distribution. Comparison of the mean intensity and standard deviation range of each panel revealed that while
normal panels displayed values similar to the standard deviation range of the entire panel area, defective panels displayed higher
values. This result is attributable to the increased intensity distribution range caused by the hot spots present on defective
panels, resulting in a larger standard deviation.
3. DEVELOPMENT AND EVALUATION OF THE
FAULT DIAGNOSIS ALGORITHM
From the analysis of intensity-related statistics, we can derive a method for assigning a reference value applicable to the fault
diagnosis algorithm. For the purpose of this study, the mean intensity and standard deviation of each panel were selected as
parameters for fault diagnosis of PV panels. First, any panel with a mean intensity that deviates from the sample standard
deviation of the sample mean intensities of the adjacent sample panels can be defined as a candidate defective panel. Here,
adjacent panels are selected from panels situated in the same array. If the intensities of each panel can be normalized, the
malfunction panel can be detected by analysing the intensities of the entire panel. However, for the normalization of the
intensities, parameters that can grasp the geometrical relationship between the target and the sensor are required.
Since these parameters were not obtained in this study, a comparison method between adjacent panels panels located in
the same array row was applied to detect malfunctioning panels with only sample images. As such, if the standard deviation is
calculated after dividing the number of the sample panels, the reference value for diagnosing a defective panel can be adjusted
for each array, and candidate defective panels can be detected efficiently. In this study, this is defined as the CMI criterion for
mean intensities. Next, any panel with a 1σ intensity range larger than the standard deviation average range of samples
standard deviations of all the adjacent panels can be classified
This contribution has been peer-reviewed. https:doi.org10.5194isprs-archives-XLII-2-W6-179-2017 | © Authors 2017. CC BY 4.0 License.
180
�
�
= �
1
− 1 �
1 2
+ �
2
− 1 �
2 2
+ ⋯ + �
�
− 1 �
� 2
�
1
+ �
2
+ ⋯ + �
�
− �
as a candidate defective panel. In this paper, this is defined as the CSD criterion for standard deviations. Here, the average
standard deviation of all the adjacent panels was calculated by assigning weight to the number of pixels using Eq. 1, in order
to reflect the different number of pixels in each panel ROI.
1 where,
S
w
= weighted average of standard deviations
n
= number of pixels of each panel
S
x
= standard deviation of each panel
k
= number of panels. The conditions for defective panel detection were set as follows:
i a mean panel intensity larger than the CMI; ii the sum of the average of standard deviations and mean intensity of the
target panel larger than the sum of the CMI and mean intensity. These detection conditions were selected because the purpose of
this study is to detect defective panels under the assumption that panels with hot spots are defective panels. Consequently, a
panel with a mean intensity lower than the CMI range on the image is considered a normal panel. This criterion setting
method, which uses 1σ range, is an empirical method derived from thermal infrared image analysis of solar power plants.
Since this method based on the empirical characteristics of intensity, various sample image analysis should be added to
improve the integrity of the algorithm. Figure 3 describes flow chart of fault diagnosis algorithm.
Figure 3. Flow chart of fault diagnosis algorithm Figure 4 presents results of diagnosing defective panels using
the proposed algorithm. Red and green squares represent the mean intensities of normal panels and panels diagnosed as
defective, respectively. Red lines represent the 1σ range of a given panel. Blue + and red × represent the CMI and CSD
ranges, respectively. In order to show whether a panel classified as a defective panel met the CSD condition, the CSD range of
the panel concerned was marked as a blue ×. Figure 4 shows that a total of seven panels were recognized as defective,
although the actual number of defective panels with hot spots on the image was confirmed as five. Hence, the algorithm was
found to have limitations in fault diagnosis by classifying panels without hot spots, such as panels 17 and 50, as defective panels.
10 20
30 40
50 60
140 150
160 170
180 190
200 210
220 230
PV Index In
te n
s it
y Mean Normal
Mean Defective 1
range of each panel
CMI CSD
Figure 4. Result of fault diagnosis algorithm application We performed additional algorithm implementation on two
images for the purpose of evaluating the proposed algorithm. Figure 5 a and b present the results of target panel labelling
after converting two thermal infrared images of 640×480 pixel size into grayscale images. The images were retrieved from the
website of the paul kitawa company, as the example image presented earlier. Drawing on our previous study Kim et al.,
2016b, in which panel area ROIs were generated from these images, we were able to evaluate the results of algorithm
implementation using the images and ROI data. There are three defective panels 11, 12, and 13 on image a and five defective
panels 4, 32, 59, 60, and 65 on image b. Figure 5 c and d present the results of panel area extraction.
Figure 5. a and b Additional test images for evaluation of the fault diagnosis algorithm www.kitawa.de, c and d Results
of panel area extraction Figure 6 and 7 represent the results of algorithm
implementation to the images in Figure 5 a and b, respectively. The results of algorithm implementation on the
two images were consistent with the defective panels recognized by visual inspection, and no diagnosis error of
recognizing a defective panel and a normal panel occurred. In Figure 6, the mean intensity of each panel was found to increase
exponentially as the row in the same array advanced. Likewise, the mean intensities of normal panels generally showed a
continuous pattern well-aligned with the adjacent panels in terms of mean intensity. These characteristics allowed us to
conclude that it was adequate to select the mean intensity of each panel as the main parameter for fault diagnosis in the
proposed algorithm. Table 1 outlines the results of performance evaluation using a confusion matrix on the basis of the number
of panels inspected. The proposed algorithm recognized all defective panels displayed on the images. As presented in Table
1, FN is 0 in all three images, i.e., completeness is 100, which means that the proposed algorithm recognized all defective
This contribution has been peer-reviewed. https:doi.org10.5194isprs-archives-XLII-2-W6-179-2017 | © Authors 2017. CC BY 4.0 License.
181
panels displayed on the images. As such, this algorithm has met the purpose of detecting defective panels. However, the
algorithm achieved an accuracy of approximately 97 in Fig. 1, which implies that a normal panel might be recognized as a
defective panel. Therefore, it is considered necessary to improve the accuracy of the algorithm by setting out additional
conditions in a future study.
10 20
30 40
50 60
100 150
200 250
PV Index In
te n
s it
y Mean Normal
Mean Defective 1
range of each panel
CMI CSD
Figure 6. Result of algorithm implementation to the image in Figure 5 a
10 20
30 40
50 60
70 140
150 160
170 180
190 200
210 220
230 240
PV Index In
te n
s it
y Mean Normal
Mean Defective 1
range of each panel
CMI CSD
Figure 7. Result of algorithm implementation to the image in Figure 5 b
Table 1. Comparison of fault detection accuracies
Image TP
FN FP
TN Completeness
Accuracy [No. of Panels]
[]
Fig. 1 5
2 61
100 97.06
Fig. 5 a 3
63 100
100 Fig. 5 b
5 65
100 100
4. CONCLUSIONS
